"1,046 requests per second per watt"
"1,046 requests per Joule"?
Intel has hit back at upstart server rival Calxeda, which claimed its ARM-powered servers could out do Chipzilla's machines. Calxeda pitted its ECX-1000 processors against Intel's Xeons on the Apache Bench benchmark test and boasted that its ARM chips had a sizeable advantage. Not so, says Intel, and it has the benchmark tests …
"1,046 requests per second per watt"
"1,046 requests per Joule"?
A datapoint for an actual server is the Boston Viridis server which has 48 Calxeda nodes (192 cores), 192GB of RAM and 24 disks. Total power (including disks, fans and power supply) is less than 300W, or just 6.25W per node (a lot less than the random guess of 11W per node!). So you end up with about 880 requests/J, still 29% better than the Ivy Bridge server.
But remember the Viridis has 24 times more DRAM, 24 times more disks, 100 times the IO bandwidth (2x 10Gbe switches per node) and does 7.5 times the number of requests/second. Therefore to make the comparison fair, you'd have to kit out 8 Ivy Bridge servers with the same amount of memory, disks and IO bandwidth, and then measure total power consumption.
Finally things will become even more interesting when Calxeda moves to a 28nm process and switches from Cortex-A9 to Cortex-A15.
Based on http://www.zdnet.com/boston-viridis-192-core-server-consumes-only-300-watts-of-datacenter-power-7000001654/ it appears that 300W does /not/ include the HDDs.
Clearly using SSDs will add the least power (but the most cost), while high performance spinning rust will use more. Looking at the spec of Seagate's Savvio 15k 2.5" disks: operating power is almost 8W.
So with 24 of them that's almost another 200W.
Still low compared to x86, but getting a long way from the 300W headline.
Be very careful when looking at manufacturer's specifications in the area they are claiming leadership: their information is spinning more than the HDDs they're not including.
Benchmark numbers are just there to reel in the CxO's - they don't have to relate to anything in real life do they?
It would be nice for vendor benchmarks to mean something, but invariably the results always contain a few * notes to the extent that these results are only possible with a previously unseen configuration that will be many times more expensive than the systems that they are actually trying to sell you.
Personally, I long for the days when you could place a marketing persons head on a stake outside your workplace to keep the evil away....
"What Calxeda and any other server chip upstart (like Applied Micro or Tilera) really need to do is run some industry standard SPEC tests on their machines. SPEC CPU2006, jEnterprise 2010, and SPECpower_ssj2008 seem like good places to start."
I would argue that synthetic benchmarks like this are the worst possible sort for realistically assessing the performance of hardware. Apachebench, MySQL bench, or (my favourite) brute force massively parallel timed kernel compile test are much more meaningful when it comes to assessing real world performance. Particularly, it is very difficult to fool these tests with applications of "smoke and mirrors" (e.g. see: http://www.altechnative.net/2012/08/04/virtual-performance-part-1-vmware/ ).
The other issue that is ignored here is the aspect of the size of the machine relative to requirements. If you need a low power low performance system. So really, what you should test is the peak performance of an ARM node and apply the same load to the x86 box. x86 box will be spinning idle for a fraction of the time, but it's overall power consumption will be greater. So if you only need that much capacity, ARM comes out ahead. The only way you can combat this on x86 is consolidation by virtualization, but that loses you at least 21%+ of performance before you've even done anything (see the link above) which levels out the playing field a bit.
But overall, I agree - there is an issue with benchmarking approach here because the test does not produce a reasonable like-for-like comparison.
SPEC is hardly a synthetic CPU benchmark. Look it up, it contains a lot of software that people use day-to-day.
+1 regarding the usage scenario, though it might be calxeda's fault for starting down that particular track.
While it would be great if arm won on ppw in all scenarios, would anyone be surprised if a power station provides more efficient power generation than your car engine?
Would anyone surprised that big iron is more efficient if you crank up the workload? There's a good reason that we don't put xeons into tablets. The point is what happens when you only need a little? You could virtualise with vmware (++license costs & resources) or some native linux vm mechanism (still adds resource overheads and management costs).
There's a market for both here and arm is only just getting going.
ARM is looking viable.
So why doesn't Intel have its cake AND eat it?
Intel has their own foundry
Intel has some of the best production, marketing and process engineering
Intel has an ARM licence, and certainly after the ARM chip sale to Marvell kept one or two Comms SoC.
So why isn't Intel making an SoC with four ARM cores, massive L1 cache, all the I/O (disk I/O) & memory interfaces... Just add RAM, ethernet magnetics and power?
It's not an either or. With MS promotion of ARM, likely Apple move to ARM for Mac Air and later all Macs and Intel promotion of Linux what is their problem?
Does Intel have an ARM licence for Cortex cores? That's not to say they couldn't get one, but I don't think they're likely to. It's hard to sell your own-design chips (for which I'd imagine they would get a much greater margin than an ARM design) if you don't even seem to have confidence in their capabilities yourself.
Intel have, in general, lost the plot.
Can you name two non-x86 technical or commercial successes from Intel in the last decade or two? Too tricky? How about one success then?
Now, who can find a dozen or so massively-promoted but ultimately unsuccessful Intel developments in the same period.
All that money to burn, and what do they come up with? i860? i960? iaPX432? i2o? IA64? uBiquitous iWiMax? iDunno any more, do you?
I do know they used some of that cash pile to bribe Dell for five years or so to ensure Dell stayed loyal to Intel . Did you know that?
Intel used to have some decent ARM-based networking chips, but (a) they mostly sold off their ARM business (2) the ARM ecosystem has moved on massively and Intel haven't.
Meanwhile, Intel bought the companies who brought you VxWorks and Simics. Not to mention McAfee. Maybe they'll have more luck there. After all, people won't be worrying about the future of VxWorks on non-x86 targets, will they...
They have an ARM licence yes.
Intel only really got into ARM as a result of a lawsuit with DEC, they used them to replace their own RISC range.
They didn't sell the IXP and IOP range to Marvell when they sold their ARM business.
Are MS promoting ARM? Really really?
MS are putting a crippled OS (no Office, no domain membership) on ARM for no reason except that they can.
MS are locking down the ARM platform (cryptographically secure boot, so the boxes can run Windows only) for no good reason except that they can.
Doesn't sound like the kind of promotion most potential customers (or even most non Wintel-dependent system builders) would choose.
Does sound like the kind of promotion a threatened company whose future was inextricably tied to the survival of Wintel business desktops, because anything other than that threatens their whole ecosystem, might choose.
"...no Office..." Really?
They know that ARM tablets can be cheaper, slimmer and run longer.
Going for ARM is just another game Microsoft are playing with Intel.